If you have Alif Ensemble or Balletto parts designed into a product, Analog Devices' purchase of Alif does not strand your models — it strands your board. Alif's neural engines are licensed Arm Ethos-U cores, and Arm's Apache-2.0 Vela compiler builds a quantized model for any Ethos-U55, U65 or U85. That means an Ensemble model can be recompiled for rival Ethos-U55 parts from Renesas, Infineon and Himax. What does not move is your board, Alif's device-locked driver pack, and any design on Alif's new Ethos-U85 parts, which have no shipping second source.
ADI signed the definitive agreement on September 9: $1.35 billion in cash at closing, up to $200 million more in contingent consideration, and a close expected before the end of calendar 2026, subject to the Hart-Scott-Rodino waiting period. The weeks between signing and close are when an Alif design-in customer has the most leverage. Use them.
What ADI Is Actually Buying for $1.35 Billion
ADI is buying a shipping line of edge-AI microcontrollers built on licensed Arm cores, and neither of its announcements names a single one of those parts. The 8-K exhibit describes "AI-native microcontrollers and fusion processors" with integrated NPUs, and says Alif's silicon is "already shipping in production, with design wins across leading consumer and industrial customers." The press release makes no roadmap commitment and names no product family. ADI CEO Vincent Roche framed the deal as "AI is moving out of the data center and into the physical world, where latency, power, and trust cannot be compromised."
An NPU (neural processing unit) is a block of silicon that runs the multiply-accumulate arithmetic of a neural network far more efficiently than the CPU next to it. Here is what Alif actually ships, from its own product pages:
- Ensemble E3: two Cortex-M55 cores (400 MHz and 160 MHz) and two Ethos-U55 NPUs, one at 256 MAC/cycle "up to 204 GOPS" and one at 128 MAC/cycle "up to 46 GOPS".
- Ensemble E7: the same pair of Ethos-U55s, plus dual Cortex-A32 cores at up to 800 MHz and up to 13.5 MB of SRAM.
- Balletto B1: a 160 MHz Cortex-M55 with one 128-MAC Ethos-U55 and a Bluetooth LE 5.3 / 802.15.4 radio.
- Ensemble E4, E6, E8: one Ethos-U85 at up to 204 GOPS alongside the U55s. Alif bills these as "the world's first microcontroller with hardware-accelerated transformer networks".
The earn-out's triggers are undisclosed. The Next Web reads the contingent $200 million as ADI wanting to see volume before paying full price. This is ADI's second AI-related acquisition of 2026, after the $1.5 billion Empower Semiconductor deal, per Yahoo Finance.
Why ADI Will Sell Two NPU Lines That Don't Share a Toolchain
ADI already sells an edge-AI microcontroller line, the MAX78000 and MAX78002. Its models cannot run on Alif's Ethos-U NPUs without being rebuilt, and Alif's cannot run on it. The MAX7800x pairs an Arm Cortex-M4 and a RISC-V core with a CNN accelerator of 64 parallel processors. The line came to ADI with Maxim, and its documentation repo is still called MaximAI_Documentation.
The toolchain is ADI's own and applies only to those two devices. The ai8x-training README says "any model that is designed to run on MAX78000/MAX78002 should use these classes". Those are ADI's custom PyTorch modules, and they come with hard limits: Conv2d kernel sizes "must be 1×1 or 3×3", the supported activations are ReLU and Abs, a MAX78000 network cannot exceed 32 layers (128 on the MAX78002), and the MAX78000 has 432 KiB of kernel memory. "TensorFlow / Keras support is deprecated." The synthesis tool then turns "the PyTorch checkpoint and the model description in the YAML format" into device-specific C code.
Alif's side is the Arm standard. Vela takes models in TOSA or TensorFlow Lite/LiteRT format and emits Ethos-U command streams; version 5.2.0 shipped September 3. PyTorch's ExecuTorch has an Ethos-U backend that targets the U55, U65 and U85 through that same compiler.
That leaves one company with two model formats, two training pipelines and two sets of customers who designed against each. Neither ADI nor Alif has said which line leads. As with Qualcomm's move on Modular, the toolchain is where silicon lock-in actually lives, so ask about the toolchain before you ask about the chip.
What Ports Off Alif, and What Doesn't
The model file ports. Your board support package, your licensed driver pack and the part numbers on your bill of materials do not. Arm says the Ethos-U85 offers "the same toolchain as previous Ethos-U generations", and Vela ships named configurations such as ethos-u55-128 and ethos-u85-256. A quantized TFLite model built for an Ensemble E3's Ethos-U55 therefore recompiles, without retraining, for:
- Renesas RA8P1: a 1 GHz Cortex-M85 with an Ethos-U55 at 500 MHz, rated at 256 GOPS.
- Infineon PSOC Edge E84: a Cortex-M55 with an Ethos-U55, plus a Cortex-M33 with Infineon's NNLite accelerator.
- Himax WiseEye2: a Cortex-M55 with an Ethos-U55.
"Recompiles" is not "drops in." An E7 carries up to 13.5 MB of SRAM; an RA8P1 carries 2 MB. A vision model sized for one may not fit the other, so plan to re-run Vela against the new part's memory map and re-measure accuracy and latency.
The code under the model is the harder move. A CMSIS Device Family Pack (DFP) is the vendor's bundle of device headers, peripheral drivers, flash algorithms and board support that your firmware is built on. Alif's DFP licence says the software "may only be used with and executed on Alif Semiconductors microcontroller, integrated circuits and SOC products." Nothing you built on it follows you to another vendor's silicon.
There is a partial hedge. Alif's Zephyr port and its Zephyr HAL are Apache-2.0, so a codebase built on Zephyr rather than on the bare DFP keeps more of its application layer portable. Nothing keeps the board portable: package, pin-out and power tree all change, which means a respin.
The Ethos-U85 Parts Have No Second Source Yet
If your design uses the E4, E6 or E8 for transformer workloads, there is no alternative part you can qualify today. That is the one place where the "your model is portable" argument runs out. Arm's Ethos-U85 IP scales from 128 to 2048 MACs with native transformer support, but IP is not silicon. Alif markets its U85 parts as the first microcontrollers of their kind, and CNX Software documents them running alongside two U55s. Ambiq's announced Ethos-U85 part, the Atomiq 110, is targeted for 2027.
Your model will still compile with Vela, just for silicon that isn't shipping yet. For a U85 design, a lifecycle commitment in writing and buffer stock are the only hedges available this year.
The Case Against Worrying
The strongest counter-argument is that ADI just agreed to pay $1.35 billion for Alif's platform and has no reason to wreck it. The company is not buying a customer list; it is buying the shipping silicon those customers designed in. If the undisclosed earn-out really is tied to volume, as The Next Web suspects, ADI is paid to push Alif parts, not bury them. Precedent points the same way: when Qualcomm bought Edge Impulse, the company told users it was "continuing support for edge hardware from our wide partner base, including MCUs, CPUs, GPUs, and NPUs."
ADI's discontinuance practice is also not abrupt. The JEDEC JESD48C standard, since superseded by J-STD-048, required suppliers to allow at least 6 months from notice to place final orders and 12 months for final shipments. ADI's own PDN 16_0062 set a last-time-buy date 12 months after publication and a last-ship date 24 months after.
All of that is fair, and none of it is a commitment about your part. The realistic risk is not that an Ensemble SKU vanishes next quarter. It is that the roadmap you designed against quietly becomes a second priority behind the other NPU line, or that low-volume SKUs get rationalised once one company owns both. Acquirers' continuity promises have a short shelf life; a clause in your supply agreement does not.
What to Get in Writing Before the Deal Closes
Your leverage peaks before close, and ADI's last deal closed fast. Empower was announced May 19 and completed July 7, 49 days later. Run the same 49 days from September 9 and you land on October 28. That is arithmetic, not a forecast, since ADI only guides to year-end. As with Bending Spoons and Miro, assume the window is shorter than the guidance.
This Week:
- List every Alif part number in shipping and in-development products, with annual volume, product owner, and which NPU each design actually uses. Put it on one page.
- Split that list. Ethos-U55-only designs have a second-source path. E4/E6/E8 designs running transformer models do not. Flag the second group to the product owner by name.
- Confirm you hold the model source, meaning the quantized
.tfliteor TOSA file and the training pipeline, not just the Vela-compiled binary your firmware links. If only a contractor has them, fix that first.
This Month:
- Run a port test. Recompile one production model with Vela for an RA8P1, PSOC Edge E84 or WiseEye2 evaluation kit, then measure accuracy and latency against the Ensemble baseline. Record the engineer-days it took. That number is your real switching cost.
- Mark your DFP exposure. Tag every firmware module as built on Alif's DFP (rewrite on a move) or on the Apache-2.0 Zephyr HAL (partly portable). Estimate the board-support rewrite in weeks.
- If you are a MAX7800x customer, do the reverse exercise. An ai8x-trained model is not portable to Ethos-U without retraining, so know which products would need it.
Before Close:
- Send your Alif account team a written request, addressed to ADI as successor, asking for: lifecycle status per part number; a commitment to a last-time-buy window of at least 12 months, which matches the window in ADI's own PDN 16_0062; and confirmation that the DFP and SDK licence terms carry over unchanged.
- Ask for the JESD48C §3.4.1 options up front: that standard obliged suppliers to negotiate "wafer or die banks" and technology transfer on request. Ask while you are still a customer they are courting, not a line item they are rationalising.
- For U85 designs, size a buffer. Without a second source, inventory is the hedge. As with d-Matrix's "any hardware" promise, get the roadmap answer on paper, or assume you do not have one.
The Bottom Line
The Cortex-M era taught embedded teams that a standard CPU core makes code portable while vendor peripherals and HALs keep you in place. The Ethos-U repeats that lesson one layer up. Arm made the NPU a standard, so your model now travels well. The lock-in moved down into the driver pack, the board and the part number, which is exactly the layer a change of control puts on the table, as it did when Visa bought BioCatch and its bank customers had reason to ask for neutrality in writing.
Arm made your model portable. Nobody made your board portable. Get the part numbers in writing.
Continue Reading
- Stilla Promised Continuity. Meta Bought It for WhatsApp.
- d-Matrix Bought Wallaroo. Get 'Any Hardware' in Writing.
- Qualcomm Spent $4B to Break Nvidia's Lock on Enterprise AI
- AMD Bought Taalas. Now Name the Model You'd Freeze.
- Nvidia Buys the Bridge to Trainium. Go Grep Your Imports.
- Bending Spoons Closed Airtable in 31 Days. Miro Says Q4.